Spaces:
Paused
Paused
File size: 15,886 Bytes
07fb183 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 | """Canonical CUDA Driver API wrapper.
This module owns device/context/module/memory/launch mechanics only. It has no
Playground, PAN, Gate-IR or scientific semantics.
"""
from __future__ import annotations
import ctypes
import math
import os
from collections.abc import Sequence
from dataclasses import dataclass
from pathlib import Path
from typing import Any
from .errors import CudaDriverError, CudaRuntimeUnavailable
from .memory import DeviceAllocation
from .preflight import CooperativePreflight, validate_cuda_block_size
# CUDA Driver API attribute IDs from cuda.h.
_CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT = 16
_CU_DEVICE_ATTRIBUTE_COOPERATIVE_LAUNCH = 95
@dataclass(frozen=True, slots=True)
class DriverModule:
"""Opaque handles returned by CUDA Driver API loading."""
context: ctypes.c_void_p
module: ctypes.c_void_p
function: ctypes.c_void_p
device_ordinal: int
class CudaDriver:
"""Minimal ctypes wrapper for CUDA-1 loading, memory and kernel launch."""
def __init__(self, library: str | None = None) -> None:
candidates = (
[library]
if library
else (["nvcuda.dll"] if os.name == "nt" else ["libcuda.so.1", "libcuda.so"])
)
loader = (
getattr(ctypes, "WinDLL", ctypes.CDLL) if os.name == "nt" else ctypes.CDLL
)
last_error: OSError | None = None
for candidate in candidates:
try:
self._lib = loader(candidate)
break
except OSError as exc:
last_error = exc
else:
names = ", ".join(candidates)
raise CudaRuntimeUnavailable(
f"CUDA driver library not available; tried: {names}"
) from last_error
self._bind()
def _bind(self) -> None:
lib = self._lib
lib.cuInit.argtypes = [ctypes.c_uint]
lib.cuInit.restype = ctypes.c_int
lib.cuDeviceGet.argtypes = [ctypes.POINTER(ctypes.c_int), ctypes.c_int]
lib.cuDeviceGet.restype = ctypes.c_int
lib.cuDeviceGetAttribute.argtypes = [
ctypes.POINTER(ctypes.c_int),
ctypes.c_int,
ctypes.c_int,
]
lib.cuDeviceGetAttribute.restype = ctypes.c_int
lib.cuCtxCreate_v2.argtypes = [
ctypes.POINTER(ctypes.c_void_p),
ctypes.c_uint,
ctypes.c_int,
]
lib.cuCtxCreate_v2.restype = ctypes.c_int
lib.cuCtxDestroy_v2.argtypes = [ctypes.c_void_p]
lib.cuCtxDestroy_v2.restype = ctypes.c_int
lib.cuModuleLoad.argtypes = [
ctypes.POINTER(ctypes.c_void_p),
ctypes.c_char_p,
]
lib.cuModuleLoad.restype = ctypes.c_int
lib.cuModuleUnload.argtypes = [ctypes.c_void_p]
lib.cuModuleUnload.restype = ctypes.c_int
lib.cuModuleGetFunction.argtypes = [
ctypes.POINTER(ctypes.c_void_p),
ctypes.c_void_p,
ctypes.c_char_p,
]
lib.cuModuleGetFunction.restype = ctypes.c_int
lib.cuOccupancyMaxActiveBlocksPerMultiprocessor.argtypes = [
ctypes.POINTER(ctypes.c_int),
ctypes.c_void_p,
ctypes.c_int,
ctypes.c_size_t,
]
lib.cuOccupancyMaxActiveBlocksPerMultiprocessor.restype = ctypes.c_int
lib.cuMemAlloc_v2.argtypes = [
ctypes.POINTER(ctypes.c_uint64),
ctypes.c_size_t,
]
lib.cuMemAlloc_v2.restype = ctypes.c_int
lib.cuMemFree_v2.argtypes = [ctypes.c_uint64]
lib.cuMemFree_v2.restype = ctypes.c_int
lib.cuMemGetInfo_v2.argtypes = [
ctypes.POINTER(ctypes.c_size_t),
ctypes.POINTER(ctypes.c_size_t),
]
lib.cuMemGetInfo_v2.restype = ctypes.c_int
lib.cuMemcpyHtoD_v2.argtypes = [
ctypes.c_uint64,
ctypes.c_void_p,
ctypes.c_size_t,
]
lib.cuMemcpyHtoD_v2.restype = ctypes.c_int
lib.cuMemcpyDtoH_v2.argtypes = [
ctypes.c_void_p,
ctypes.c_uint64,
ctypes.c_size_t,
]
lib.cuMemcpyDtoH_v2.restype = ctypes.c_int
lib.cuLaunchKernel.argtypes = [
ctypes.c_void_p,
ctypes.c_uint,
ctypes.c_uint,
ctypes.c_uint,
ctypes.c_uint,
ctypes.c_uint,
ctypes.c_uint,
ctypes.c_uint,
ctypes.c_void_p,
ctypes.POINTER(ctypes.c_void_p),
ctypes.POINTER(ctypes.c_void_p),
]
lib.cuLaunchKernel.restype = ctypes.c_int
lib.cuCtxSynchronize.argtypes = []
lib.cuCtxSynchronize.restype = ctypes.c_int
@staticmethod
def _check(code: int, call: str) -> None:
if code != 0:
raise CudaDriverError(f"{call} failed with CUDA error code {code}")
def initialize(self) -> None:
self._check(self._lib.cuInit(0), "cuInit")
def _device(self, ordinal: int) -> int:
device = ctypes.c_int()
self._check(
self._lib.cuDeviceGet(ctypes.byref(device), ordinal),
"cuDeviceGet",
)
return int(device.value)
def device_attribute(self, ordinal: int, attribute: int) -> int:
device = self._device(ordinal)
value = ctypes.c_int()
self._check(
self._lib.cuDeviceGetAttribute(
ctypes.byref(value),
attribute,
device,
),
"cuDeviceGetAttribute",
)
return int(value.value)
def alloc_device(self, size_bytes: int) -> DeviceAllocation:
"""Allocate device memory owned by the current CUDA context."""
if size_bytes <= 0:
raise ValueError("size_bytes must be positive")
ptr = ctypes.c_uint64()
self._check(
self._lib.cuMemAlloc_v2(ctypes.byref(ptr), size_bytes),
"cuMemAlloc_v2",
)
return DeviceAllocation(ptr=int(ptr.value), size_bytes=size_bytes)
def free_device(self, allocation: DeviceAllocation) -> None:
"""Release one device allocation."""
if allocation.ptr:
self._check(
self._lib.cuMemFree_v2(ctypes.c_uint64(allocation.ptr)),
"cuMemFree_v2",
)
def memory_info(self) -> tuple[int, int]:
"""Return free and total bytes for the current CUDA context."""
free_bytes = ctypes.c_size_t()
total_bytes = ctypes.c_size_t()
self._check(
self._lib.cuMemGetInfo_v2(
ctypes.byref(free_bytes),
ctypes.byref(total_bytes),
),
"cuMemGetInfo_v2",
)
return int(free_bytes.value), int(total_bytes.value)
def copy_host_to_device(
self,
allocation: DeviceAllocation,
source: ctypes.Array[Any],
*,
size_bytes: int | None = None,
) -> None:
"""Copy a ctypes-backed host buffer to device memory."""
count = allocation.size_bytes if size_bytes is None else size_bytes
host_bytes = ctypes.sizeof(source)
if count < 0 or count > allocation.size_bytes:
raise ValueError("host-to-device copy exceeds device allocation")
if count > host_bytes:
raise ValueError(
f"host-to-device copy exceeds host buffer: {count} > {host_bytes}"
)
self._check(
self._lib.cuMemcpyHtoD_v2(
ctypes.c_uint64(allocation.ptr),
ctypes.cast(source, ctypes.c_void_p),
count,
),
"cuMemcpyHtoD_v2",
)
def copy_device_to_host(
self,
destination: ctypes.Array[Any],
allocation: DeviceAllocation,
*,
size_bytes: int | None = None,
) -> None:
"""Copy device memory into a ctypes-backed host buffer."""
count = allocation.size_bytes if size_bytes is None else size_bytes
host_bytes = ctypes.sizeof(destination)
if count < 0 or count > allocation.size_bytes:
raise ValueError("device-to-host copy exceeds device allocation")
if count > host_bytes:
raise ValueError(
f"device-to-host copy exceeds host buffer: {count} > {host_bytes}"
)
self._check(
self._lib.cuMemcpyDtoH_v2(
ctypes.cast(destination, ctypes.c_void_p),
ctypes.c_uint64(allocation.ptr),
count,
),
"cuMemcpyDtoH_v2",
)
def launch_kernel(
self,
loaded: DriverModule,
*,
grid: tuple[int, int, int],
block: tuple[int, int, int],
arguments: Sequence[object],
dynamic_shared_bytes: int = 0,
) -> None:
"""Launch a loaded kernel with CUDA Driver API argument packing."""
if min(*grid, *block) <= 0:
raise ValueError("grid and block dimensions must be positive")
storage = list(arguments)
kernel_params = (ctypes.c_void_p * len(storage))()
for index, argument in enumerate(storage):
kernel_params[index] = ctypes.cast(
ctypes.byref(argument), ctypes.c_void_p # type: ignore[arg-type]
)
self._check(
self._lib.cuLaunchKernel(
loaded.function,
grid[0],
grid[1],
grid[2],
block[0],
block[1],
block[2],
dynamic_shared_bytes,
None,
kernel_params,
None,
),
"cuLaunchKernel",
)
def synchronize(self) -> None:
"""Synchronize the current CUDA context."""
self._check(self._lib.cuCtxSynchronize(), "cuCtxSynchronize")
def launch_cooperative(
self,
loaded: DriverModule,
*,
n_neurons: int,
block_size: int,
arguments: Sequence[object],
) -> CooperativePreflight:
"""Launch a resident 1D grid only after checking actual kernel occupancy."""
preflight = self.cooperative_preflight(
loaded, n_neurons=n_neurons, block_size=block_size
)
if n_neurons < 1 or not preflight.launch_fits:
raise ValueError(
"cooperative grid exceeds the device/kernel resident capacity"
)
launch = self._lib.cuLaunchCooperativeKernel
launch.argtypes = [
ctypes.c_void_p,
ctypes.c_uint,
ctypes.c_uint,
ctypes.c_uint,
ctypes.c_uint,
ctypes.c_uint,
ctypes.c_uint,
ctypes.c_uint,
ctypes.c_void_p,
ctypes.POINTER(ctypes.c_void_p),
]
launch.restype = ctypes.c_int
storage = list(arguments)
parameters = (ctypes.c_void_p * len(storage))()
for index, argument in enumerate(storage):
parameters[index] = ctypes.cast(
ctypes.byref(argument), ctypes.c_void_p # type: ignore[arg-type]
)
self._check(
launch(
loaded.function,
preflight.required_blocks,
1,
1,
block_size,
1,
1,
0,
None,
parameters,
),
"cuLaunchCooperativeKernel",
)
return preflight
def load_cubin(
self,
cubin_path: Path,
*,
kernel_name: str = "pan_gate_kernel",
device_ordinal: int = 0,
) -> DriverModule:
"""Load a ptxas-generated cubin and resolve one kernel symbol."""
self.initialize()
device = self._device(device_ordinal)
context = ctypes.c_void_p()
self._check(
self._lib.cuCtxCreate_v2(ctypes.byref(context), 0, device),
"cuCtxCreate_v2",
)
module = ctypes.c_void_p()
try:
self._check(
self._lib.cuModuleLoad(
ctypes.byref(module),
str(cubin_path).encode("utf-8"),
),
"cuModuleLoad",
)
function = ctypes.c_void_p()
self._check(
self._lib.cuModuleGetFunction(
ctypes.byref(function),
module,
kernel_name.encode("ascii"),
),
"cuModuleGetFunction",
)
except Exception:
if module.value:
self._lib.cuModuleUnload(module)
self._lib.cuCtxDestroy_v2(context)
raise
return DriverModule(
context=context,
module=module,
function=function,
device_ordinal=device_ordinal,
)
def kernel(self, loaded: DriverModule, name: str) -> DriverModule:
"""Borrow another function in the same owned module/context; do not unload it separately."""
if not name or not name.isascii() or not name.replace("_", "").isalnum():
raise ValueError("invalid CUDA kernel symbol")
function = ctypes.c_void_p()
self._check(
self._lib.cuModuleGetFunction(
ctypes.byref(function), loaded.module, name.encode("ascii")
),
"cuModuleGetFunction",
)
return DriverModule(
loaded.context, loaded.module, function, loaded.device_ordinal
)
def unload(self, loaded: DriverModule) -> None:
module_error: CudaDriverError | None = None
if loaded.module.value:
try:
self._check(
self._lib.cuModuleUnload(loaded.module),
"cuModuleUnload",
)
except CudaDriverError as exc:
module_error = exc
if loaded.context.value:
try:
self._check(
self._lib.cuCtxDestroy_v2(loaded.context),
"cuCtxDestroy_v2",
)
except CudaDriverError:
if module_error is None:
raise
if module_error is not None:
raise module_error
def cooperative_preflight(
self,
loaded: DriverModule,
*,
n_neurons: int,
block_size: int = 128,
dynamic_shared_bytes: int = 0,
) -> CooperativePreflight:
"""Calculate whether the full cooperative grid can be resident."""
validate_cuda_block_size(block_size)
cooperative = bool(
self.device_attribute(
loaded.device_ordinal,
_CU_DEVICE_ATTRIBUTE_COOPERATIVE_LAUNCH,
)
)
sms = self.device_attribute(
loaded.device_ordinal,
_CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT,
)
active = ctypes.c_int()
self._check(
self._lib.cuOccupancyMaxActiveBlocksPerMultiprocessor(
ctypes.byref(active),
loaded.function,
block_size,
dynamic_shared_bytes,
),
"cuOccupancyMaxActiveBlocksPerMultiprocessor",
)
required = math.ceil(n_neurons / block_size)
capacity = int(active.value) * sms
return CooperativePreflight(
cooperative_launch=cooperative,
multiprocessor_count=sms,
active_blocks_per_sm=int(active.value),
block_size=block_size,
required_blocks=required,
resident_block_capacity=capacity,
launch_fits=cooperative and required <= capacity,
n_neurons=n_neurons,
)
|